Inspiration
Food donation can help reduce food waste and support communities, but reviewing surplus food donations requires more than simply knowing what food is available. Donation teams need relevant information about the food, its quantity, preparation, storage history and packaging condition before making decisions.
We wanted to explore how AI could make this evidence-review process more organised, helping teams identify missing information and bring important details to human reviewers.
What it does
FoodRescue Sentinel is an AI-powered evidence review assistant for surplus-food donations.
It is designed to:
- Extract structured information from natural-language donor messages.
- Distinguish confirmed, inferred and unknown information.
- Identify missing evidence and request clarification.
- Evaluate available information against verified, source-backed rules.
- Generate structured review reports for human decision-making.
- Maintain case records and human review decisions.
FoodRescue Sentinel supports the review process; it does not independently certify food as safe or authorise distribution.
How we built it
We developed the backend using Python and FastAPI, with SQLite for case management, message history and review records.
The system uses a structured data model to organise donation information and a rules engine to evaluate available evidence. We also implemented API endpoints for submitting donor messages, retrieving cases, checking rules and recording human review decisions.
Our planned AI integration uses NVIDIA Nemotron through Nebius infrastructure to interpret donor messages and help generate structured assessments.
Challenges we faced
One of our main challenges was handling incomplete information. A donor's message may not contain every detail needed for a meaningful review. We therefore designed the system to distinguish confirmed information from inferred or unknown details rather than treating missing information as fact.
Another challenge was ensuring that AI-generated information would not be confused with verified food-safety guidance. We addressed this through a separate rules engine and a human-review workflow.
We also faced infrastructure and integration challenges while setting up access to Nebius Token Factory for the hackathon.
What we learned
Throughout this project, we learned how to connect AI workflows with backend APIs, structure information for consistent evaluation, manage case data and design systems that keep human decisions involved in important processes.
Most importantly, we learned that useful AI systems should not simply generate answers. They should make uncertainty visible, provide traceable information and help people make better-informed decisions.
Log in or sign up for Devpost to join the conversation.